对于小规模的数据集我们可以将其一次性读入内存（CPU）中然后再分batch让GPU去训练，只要简单地使用fit函数即可；然而当数据集规模超大时，一次性读入所有数据会使得内存溢出（与GPU无关，GPU是显存，batch_size过大才会导致显存爆炸），我们需要使用fit_generator来训练
如下，
fit_generator(generator, steps_per_epoch=None,
epochs=1, verbose=1,
callbacks=None, validation_data=None,
validation_steps=None, class_weight=None,
max_queue_size=10, workers=1,
use_multiprocessing=False, shuffle=True, initial_epoch=0)
它具有很多参数，其中比较重要的参数是generator，我们可以使用python的生成器，也可以使用keras.utils中的Sequence，个人比较推荐后者，它需要重写两个函数：
- len()
- getitem()
其实也很简单，以下举个简单的例子你便能学会写自己的数据自动生成器。
代码如下，接下来一个个函数解释。
import numpy
import matplotlib.image as mpimg # read image
from keras.utils import Sequence, to_categorical
class DataGenerator(Sequence):
def __init__(self, files, batch_size=1, shuffle=True):
"""
# Arguments
---
files: filename.
batch_size: . """
self.batch_size = batch_size
self.files = files
self.indexes = numpy.arange(len(self.files))
self.shuffle = shuffle
def __len__(self):
"""return: steps num of one epoch. """
return len(self.files) // self.batch_size
def __getitem__(self, index):
"""Gets the `index-th` batch.
---
# Arguments
index: position of the batch in the Sequence.
# Returns
A batch data. """
# get batch data inds.
batch_inds = self.indexes[index *
self.batch_size:(index+1)*self.batch_size]
# get batch data file name.
batch_files = [self.files[k] for k in batch_inds]
# read batch data
batch_images, batch_labels = self._read_data(batch_files)
return batch_images, batch_labels
def on_epoch_end(self):
"""shuffle data after one epoch. """
if self.shuffle == True:
numpy.random.shuffle(self.indexes)
def _read_data(self, batch_files):
"""Read a batch data.
---
# Arguments
batch_files: the file of batch data.
# Returns
images: (batch_size, (image_shape)).
labels: (batch_size, (label_shape)). """
images = []
labels = []
for file in batch_files:
image = mpimg.imread('data/Images/'+file+'.jpg')
images.append(image)
lable = numpy.loadtxt('data/labels/'+file+'.arr', dtype=int)
labels.append(lable) # to one hot
return numpy.array(images), numpy.array(labels)
init
首先定义自己的数据生成器 DataGenerator 继承Sequence类，初始化参数
class DataGenerator(Sequence):
def __init__(self, files, batch_size=1, shuffle=True):
"""
# Arguments
---
files: filename.
batch_size: . """
self.batch_size = batch_size
self.files = files
self.indexes = numpy.arange(len(self.files))
self.shuffle = shuffle
- files：所有训练样本的文件路径
- batch_size：一个batch的大小
例如我做图像分割任务，图片大概是这样的
label是像素标签矩阵，与上面的每张图片对应
因此我只需要除了文件后缀之外的字符串即可，当读取对应的图片或label时我再在后面添加对应的后缀。所以传入的 files 大概是这样的：
并且创建一个index变量用来划分数据集。
len
len函数计算每一个epoch的步长，即每个epoch有多少个batch要训练
def __len__(self):
"""return: steps num of one epoch. """
return len(self.files) // self.batch_size
getitem函数要求返回一个batch的data，包括输入数据和标签，参数是index（训练的时候会自动调用），代表第index个batch，通过这个index我们去计算这个batch的数据的下标（batch_inds），然后提取出它对应的文件名字（batch_files），把这个batch_files传入给_read_data函数让它取读取这个batch的image和label
def __getitem__(self, index):
"""Gets the `index-th` batch.
---
# Arguments
index: position of the batch in the Sequence.
# Returns
A batch data. """
# get batch data inds.
batch_inds = self.indexes[index *
self.batch_size:(index+1)*self.batch_size]
# get batch data file name.
batch_files = [self.files[k] for k in batch_inds]
# read batch data
batch_images, batch_labels = self._read_data(batch_files)
return batch_images, batch_labels
_read_data
这个函数就是根据文件名去读取自己对应任务的数据了。
def _read_data(self, batch_files):
"""Read a batch data.
---
# Arguments
batch_files: the file of batch data.
# Returns
images: (batch_size, (image_shape)).
labels: (batch_size, (label_shape)). """
images = []
labels = []
for file in batch_files:
image = mpimg.imread('data/Images/'+file+'.jpg')
images.append(image)
lable = numpy.loadtxt('data/labels/'+file+'.arr', dtype=int)
labels.append(lable) # to one hot
return numpy.array(images), numpy.array(labels)
on_epoch_end
这个函数是每个epoch结束后调用，是否将数据集打乱。
def on_epoch_end(self):
"""shuffle data after one epoch. """
if self.shuffle == True:
numpy.random.shuffle(self.indexes)
以上即为构造数据生成器的过程，在构造模型时，可参考以下例子。构造FCN32模型，且用两个GPU来训练，此时batchsize的大小如果是1会报错，因为另一个GPU就没有训练数据了，所以batchsize至少是2，如果是单个GPU训练则不影响。
def get_generator(num_classes, batch_size=2, preprocess=True, shuffle=True, train_ratio=0.8):
"""Get data generator for training and test files.
---
# Arguments
num_classes: .
preprocess: preprocess image data.
shuffle: shuffle data after epoch.
train_ratio: split data.
# Returns
generator: data generator for fit_generator.
test_files: test model after training. """
images_files = os.listdir('data/Images')
allfiles = []
for file in images_files:
allfiles.append(file.split('.')[0]) # get image name
N = len(allfiles)
train_files = allfiles[:int(N*train_ratio)]
test_files = allfiles[int(N*train_ratio):]
generator = DataGenerator(train_files, num_classes,
batch_size, preprocess, shuffle)
return generator, test_files
num_classes = 34
# Get generator.
generator, test_files = get_generator(num_classes,
batch_size=2,
preprocess=True,
shuffle=True,
train_ratio=0.8)
# Build model.
model = FCN32.build(height=256, width=256, num_classes=num_classes)
model = multi_gpu_model(model, gpus=2) # 两个GPU, batch_size至少应该是2
model.summary()
model.compile(optimizer=Adam(0.00001),
loss='categorical_crossentropy',
metrics=['accuracy'])
# fit model.
model.fit_generator(generator, epochs=20)
